Researchers led by Peng Zhang have presented CRAFT (Contextual Requirement Adaptation with Frozen Tools), a framework that lets a robot learn where objects belong from people's corrections without retraining its skills, according to a paper on arXiv. If a person turns down where the robot put something, it has to work out whether the correction holds just for that case or stands for a wider rule, the authors write.

CRAFT records whether an action physically worked separately from whether it met the person's requirement. Where an almost identical situation has been judged before, it falls back on memory; elsewhere it combines memory, explicit rules and a learner, weighting each by how well it predicted earlier feedback.

In simulations with cameras, CRAFT asked fewer questions than language-model agents with lasting memory, but some of those agents made fewer wrong moves and succeeded more often after the requirements changed. Everything was tested in simulation, with generated requirements and simulated feedback.

Source: arXiv Robotics (cs.RO)

Background

Peng Zhang

Robotics researcher

Robotics researcher.